Nature Computational Science review maps four roles for LLMs as human proxies, each needing its own validity criteria
iyadrahwan · x · 2026-09-21
Karetnikov, Iyad Rahwan, and Davor Svetinovic published a review, "Large language models as human proxies," in Nature Computational Science, analyzing research across disciplines that explicitly or implicitly use LLMs to stand in for humans.
The paper organizes these uses into four roles:
- Believable agents
- Task agents
- Experimental subjects
- Silicon samples
Key conclusion: similarity to humans is not a single feature—each role supports a different scientific claim and requires its own validity criteria, providing a taxonomy for the growing practice of computationally approximating human behavior with LLMs.
More from AGI Musings
- AI now beats some of the best human forecasters, The Economist reports — 233C · 2026-09-21
- HF incident was mostly harmless, but proves your precautions are futile — ctjlewis · 2026-09-21
- Today's models already outsmart humans, so why talk of guarding against superintelligence? — ctjlewis · 2026-09-21
- Engineer turns down $100k remote job fearing one Claude update could kill his product — sharpeye_wnl · 2026-09-21
- 90%+ of code will be written by LLMs, yet human instruction volume keeps growing — timigod · 2026-09-21
- ICLR 2027 and beyond: is it the author or the agent who's doing the work? — MuCai7 · 2026-09-21